AI Models Cross Capability Threshold in Binary Exploitation Benchmarks

Anthropic's Frontier Red Team reported that frontier AI models including GLM-5.3 and Claude Mythos Preview achieved full control flow hijacks in binary exploitation benchmarks. This marks a notable capability shift, as previous generation models such as Claude Opus 4.6 and GLM-5.2 failed to achieve any successful hijacks on the same benchmark. This milestone demonstrates that advanced LLMs are acquiring non-trivial autonomous offensive cyber capabilities at low levels of software execution. It highlights the growing urgency of proactive AI safety testing and red teaming to assess risks before deploying powerful frontier models. Across 100 internal binary exploitation benchmark tasks, Claude Mythos Preview achieved full control flow hijacks in 6% of trials, while GLM-5.3 succeeded in 4%. Earlier models failed in 100% of trials, indicating a clear qualitative leap in vulnerability exploitation capability.